Chronic Respiratory Disease: Risk Modeling Potential and Limitations

Chronic Respiratory Diseases (CRDs), including Chronic Obstructive Pulmonary Disease (COPD) and asthma, are among the leading causes of mortality worldwide, with 545 million prevalent cases in 2017. Symptoms of noninfectious CRDs are often exacerbated by ambient air pollution and changes in temperature and humidity. This study explores a novel application of machine learning to forecast CRD risk and discuss its merits and limitations. We developed, trained, and tested a Random Forest regressor using datasets over the United States during 2000-2016, with mortality rate as the target variable. The final regressor produced R-squared values of 0.7526 and 0.7528 for cross-validation and test dataset prediction, respectively, implying that our model generalizes well out-of-sample. The selected features comprise location and temporal encoders, population density, and net primary production. The study reveals significant potential for modeling CRD risk but highlights setbacks due to the primarily noninfectious nature of CRDs, phenomena only identifiable on finer spatiotemporal scales, and data limitations. We also identify methods that may refine our approach and describe future developments that may improve CRD risk modeling.